Anthropic Economic Index: Insights from Claude 3.7 Sonnet

Source: Anthropic | Date: Mar 27, 2025 Report: Second release from the Anthropic Economic Index, analyzing usage data from Claude.ai following the launch of Claude 3.7 Sonnet.

Key Findings Summary

1. Shifts in Usage Patterns

• Increased Share: Modest rise in the share of usage for coding, education, sciences, and healthcare applications since the launch of Claude 3.7 Sonnet.
• Expected vs. Unexpected: The increase in coding was anticipated due to improved benchmarks. Increases in other categories may reflect broader AI diffusion, novel applications, or unexpected model improvements.

2. Use of "Extended Thinking" Mode

• Primary Use Case: Predominantly used for technical and creative problem-solving.
• Top Occupations by Usage:
  • Computer and Information Research Scientists (~10%)
  • Software Developers (~8%)
  • Multimedia Artists (~7%)
  • Video Game Designers (~6%)

3. Augmentation vs. Automation

• Overall Balance: Remains essentially unchanged, with augmentation comprising 57% of usage.
• Notable Shift: "Learning" interactions (asking for information/explanation) rose from ~23% to ~28%.
• By Occupational Category:
  • Highly Augmentative: Community and Social Service tasks (approaching 75% augmentation).
  • Balanced (~50/50): Production and Computer & Mathematical occupations.
  • No categories were automation-dominant.
• By Specific Occupation (Examples):
  • Highest Task Iteration (Human-AI Co-writing): Copywriters and Editors.
  • Highest Directive Behavior (Minimal Human Involvement): Translators and Interpreters.

4. New Bottom-Up Taxonomy of Usage

• Purpose: To capture use cases potentially missed by the top-down O*NET task database.
• Dataset: 630 granular categories organized in a three-level hierarchy.
• Example Categories:
  • Help resolve household plumbing, water, and maintenance issues
  • Provide guidance on battery technologies and charging systems
  • Help with time zone handling in code and databases
  • Create physics-based simulations with interactive visualization capabilities

Key Excerpts & Data Points

"Since the launch of Claude 3.7 Sonnet, we've observed a rise in the share of usage for coding, as well as educational, science, and healthcare applications."

"Tasks associated with copywriters and editors show the highest amount of task iteration, where the human and model co-write something together. By contrast, tasks associated with translators and interpreters show among the highest amounts of directive behavior—where the model completes the task with minimal human involvement."

• Data Analyzed: 1 million anonymized Claude.ai Free and Pro conversations from the 11 days post-launch.
• Methodology: Conversations mapped to 17,000 tasks in the U.S. Department of Labor's O*NET database using the privacy-preserving tool Clio.
• Datasets Released: All analysis datasets are freely available for download on Anthropic's Hugging Face page.

Methodological Notes

• Model Used for Analysis: Claude 3.7 Sonnet replaced Claude 3.5 Sonnet for classification tasks, improving accuracy.
• Data Filtering: Conversations flagged by safety classifiers were removed; no occupational relevance filter was applied, preserving more data for the bottom-up taxonomy.

Conclusion & Future Work

• The report tracks modest shifts in usage patterns and confirms the stability of the augmentation/automation balance.
• Anthropic aims to continue tracking these metrics and developing new ones as AI capabilities and applications evolve.
• Call for Researchers: Anthropic is hiring for Societal Impacts Research Scientist, Research Engineer, and Economist roles to study AI's effects on the labor market.

Appendix Highlights

• Depth of Task Usage Curve: Shows little change from the first report, with a slight potential increase in concentration towards coding.
• Largest Occupational Category Increase: Computer and mathematical occupations saw a +3% absolute increase in usage share.